The present disclosure relates to creating sets of scored paraphrases based on a structured resource and using the sets of scored paraphrases to identify and score candidate answers that correspond to a user question.
A traditional search engine produces the most accurate answers when words in a question match a passage in a corpus of documents in the same order. For example, if the traditional search engine receives a question of “Who is the president of Company ABC?”, the traditional search engine produces accurate results if the corpus includes a document with the passage “Bill Smith is the president of Company ABC.” If a passage in the corpus of documents does not closely match the words of the question in order, the traditional search engine is less likely to produce an accurate answer.
In reality, a corpus of documents may not have matching paraphrases, but rather have passages that include answers to a question such as “Bill Smith leads Company ABC”, “The CEO of Company ABC is Bill Smith”, or “Company ABC's Chairperson is Bill Smith”. Unfortunately, the traditional search engine may not detect information in these passages to determine candidate answers.